Subject: port-alpha/7826: poor network performance
To: None <gnats-bugs@gnats.netbsd.org>
From: Michael Stolarchuk <mts@alpha1.pdq.com>
List: netbsd-bugs
Date: 06/21/1999 05:50:48
>Number:         7826
>Category:       port-alpha
>Synopsis:       poor network performance
>Confidential:   no
>Severity:       serious
>Priority:       high
>Responsible:    port-alpha-maintainer (NetBSD/alpha Portmaster)
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Mon Jun 21 05:50:00 1999
>Last-Modified:
>Originator:     Michael Stolarchuk
>Organization:
mts@pdq.com
>Release:        1.3, 1.32
>Environment:
NetBSD alpha1 1.3I NetBSD 1.3I (ALPHAGW) #2: Fri Apr 9 09:05:57 EDT 1999 root@alpha1:/usr/src/sys/arch/alpha/compile/ALPHAGW alpha
NetBSD 1.3I (ALPHAGW) #2: Fri Apr  9 09:05:57 EDT 1999
    root@alpha1:/usr/src/sys/arch/alpha/compile/ALPHAGW
Digital AlphaPC 164LX 533 MHz, 531MHz
8192 byte page size, 1 processor.
real mem = 1073741824 (1949696 reserved for PROM, 1071792128 used by NetBSD)


>Description:
poor net performance due to packet reception problems, 100MHz net, 164lx's

Problem was originally noticed while doing ftp's between a group of 
eight 533 164lx alphas with 1G physical memory.  First tested with
de500's, then was able to replicate the problem with intel express
cards, and cogent cards, then with switches in place of hubs, different cables, etc.
For the 100Mhz net, peak ftp performance end-end was approxmiatly 80-100kb.

Mods to several of the drivers to modify transmission parameters resulting
in no significant increase in transmission rate.  The receiver has trouble with
6-7th received packet; typically an overrun or crc error.  That packet
loss coupled with the long retransmission timer for tcp causes the
poor performance.

The performance can be somewhat improved by lowering the tcp window.  That results
in fewer back to back packets, which can improve the rate somewhat, but still
not get the performance in the range we ought to get.

We are currently investigating whether linux has the same problems on the
same hardware.  Our first tests indicate redhad 6.0 does NOT have the
same problem. 

We have been trying to invesitage this problem for quite a number of weeks,
an would very much appreciate help in resolving the situation.

>How-To-Repeat:
	ftp between 164lx's
>Fix:
	
>Audit-Trail:
>Unformatted: